Cloudflare OS: installing and setting up
As of September 2026, compiled from our own notes and news archive; text machine-generated and editorially checked. Check versions and commands in the official documentation.
In this guide you set up Cloudflare OS: Cloudflare's open-source AI working environment in which an agent builds small apps ("gadgets") for you and performs tasks, securely isolated. You try it locally on your own computer or set it up in your own Cloudflare account. Everything follows the repository's official README (as of September 2026).
First, a quick word: what are you installing?
Cloudflare OS is not an operating system for your computer. It is a web application that runs on Cloudflare's Workers runtime. You work with it in the browser, roughly like an online office suite. Except that the agent creates a separate small app for each "document", which runs in its own sandbox.
Cloudflare explicitly calls this version (v2, released in August 2026) early access: much already works, but there are still rough edges.
What you need
- Trying it locally: Git, Node.js and the package manager pnpm. A Cloudflare account is not required: the runtime (workerd) is open source and runs on your own machine.
- Putting it online: a Cloudflare account.
- A language model: Cloudflare OS works with various major AI providers and with self-hosted models.
Option 1: try it locally
Fetch the code and start the entire environment:
git clone https://github.com/cloudflare/cloudflare-os.git
cd cloudflare-os
pnpm run-localThen open http://localhost:8787 in your browser.
This runs the full stack locally via wrangler (Cloudflare's development tool) on top of workerd. Your data goes into the .wrangler subdirectory. Note: this is intended for trying it out, not for running production.
Option 2: in your own Cloudflare account
The simplest route is Cloudflare's online deployment process:
If you want more control, for example your own gatekeepers or changes to the code, use the starter repository as a basis:
github.com/cloudflare/cloudflare-os-starter
Running on your own server with only workerd, without Cloudflare, is in principle also possible, but the documentation and tooling for that are still "coming soon".
The first steps in the environment
Log in and try a few of the example prompts from the README:
- "Make slides for my upcoming meeting with a client." This uses the built-in blueprint for presentations.
- "Make a collaborative whiteboard app." The agent then builds a new app from scratch.
- "Make a tic-tac-toe game", followed by "I am X, you are O. I have made my first move, you are."
Every app created this way is a gadget: private and isolated, but shareable with others, just like a document in an online office suite.
Connecting external services (gatekeepers)
By default, gadgets and the agent have access to nothing, not even the internet. Access to services such as GitHub or Google goes through a gatekeeper: a separate component that monitors the connection, logs everything and presents actions with consequences to you first.
For most gatekeepers, you have to create OAuth credentials once with the service itself. Each gatekeeper has its own guide in the repository (packages/gatekeeper-*/README.md). Available ones include:
- GitHub, Google, Cloudflare and Supabase.
- Notion, Confluence and Slack.
- Email Workers, Home Assistant, Spotify and ZoomInfo.
Start with one service you really need. Creating OAuth apps is cumbersome with some providers.
Checking that it works
- The web interface loads at
localhost:8787(locally) or on your own domain (in the cloud). - A simple prompt such as the tic-tac-toe game produces a working gadget.
- After connecting a gatekeeper, you can "suggest" that service to a chat or gadget (for example by pasting a link to a GitHub repository), and the agent can then work with it.
Updating
Locally, you fetch the latest code and start again:
git pull
pnpm run-localFor an installation via the starter repository, you update it according to the instructions there. Because the project is under active development, updates can change a great deal. So first make a copy of the .wrangler folder if you want to keep your local data.